Anthropic has issued a warning that its Claude AI models are being exploited by cybercriminals and state-sponsored hackers. The company's latest report, covering activity between December 2025 and August 2026, details how Claude Haiku, Sonnet, and Opus models have been used for purposes such as automating cyberattacks, conducting mass surveillance, and attempting to develop more dangerous versions of deadly viruses.
A Russian espionage group, tracked by Anthropic as GTG-20006, reportedly used AI to automate their entire cyberattack kill chain, increasing the speed of their operations. This group targeted more than 20 organisations, including embassies, think tanks, and government agencies across multiple regions.
The report also highlights five cases of users in "unsupported regions" attempting to use Claude to support biological weapons development. One instance involved a scientist trying to use Claude to write a grant application for research into the transmissibility and immune evasion properties of the chikungunya virus, which could potentially make the pathogen more dangerous. Another case in May involved a user outside the US researching adaptations of highly pathogenic avian influenza.
Furthermore, Anthropic identified six cases where users outside the US employed Claude to develop software for conventional weapons, including guidance, navigation, and control (GNC) software for flying vehicles. In Yemen, a weapons development program used Claude for GNC software, and the same team reportedly test-fired a guided rocket. In China, an actor associated with a defense industry manufacturer used Claude to draft specifications for an anti-torpedo fire control system.